An Explainable Machine Learning Framework for Real-Time Loan Default Prediction Using Customer Financial Behavior and Foreign Exchange Market Indicators

Abstract

Accurate prediction of loan defaults is paramount for financial institutions to mitigate credit risk, yet the complexity of modern machine learning models often compromises interpretability, undermining stakeholder trust and regulatory compliance. This study proposes a novel explainable machine learning framework that integrates customer financial behavior indicators with foreign exchange (FX) market variables for real-time loan default prediction. Employing a hybrid approach combining XGBoost for predictive accuracy and SHapley Additive exPlanations (SHAP) for model interpretability, the framework is evaluated on a comprehensive dataset of consumer loans. The proposed model demonstrates superior predictive performance, achieving an AUC-ROC of 0.94 and an accuracy of 91%, significantly outperforming traditional logistic regression. SHAP analysis reveals that interest rates, loan amounts, and debt-to-income ratios, alongside FX volatility, are the most critical determinants of default, with the influence of these features varying with market conditions. The framework's explainability facilitates transparent decision-making, allowing risk managers to understand the rationale behind predictions and comply with emerging regulatory requirements. This research contributes a practical, interpretable solution that bridges the gap between high-performance predictive modeling and the critical need for accountability in financial risk management.